README.TXT
资源名称:netcat.zip [点击查看]
上传用户:shhspump
上传日期:2007-01-07
资源大小:85k
文件大小:61k
源码类别:
远程控制编程
开发平台:
Visual C++
- THE DEFCON RELEASE 7/26/96
- This is a beta release of netcat for Windows 95 and Windows NT. It was
- ported from hobbit's netcat 1.0.
- This was built on NT 4.0 using Visual C++ 4.0. It has had some superficial
- testing on Windows 95 as well as NT 3.51 and 4.0.
- *Things that don't work yet*
- - None of the timer settings for delays and timeouts are implemented.
- - source routing may never work on NT.
- - the gaping security hole feature or exec function is not done.
- I am planning on cleaning up the gnarly IFDEFs for Win32 and brining this
- release back into the fold of the standard unix make if possible. NT
- pretends to be unix but it's only pretty close.
- Please report any problems to the address below.
- Weld Pond (weld@l0pht.com)
- L0pht Heavy Industries
- NOW ON TO hobbit's FULL README FILE
- Netcat 1.10
- =========== /_/
- / 0 0
- Netcat is a simple Unix utility which reads and writes data ====v====
- across network connections, using TCP or UDP protocol. W /
- It is designed to be a reliable "back-end" tool that can | | _
- be used directly or easily driven by other programs and / ___ /
- scripts. At the same time, it is a feature-rich network / / |
- debugging and exploration tool, since it can create almost (((-----)))-'
- any kind of connection you would need and has several /
- interesting built-in capabilities. Netcat, or "nc" as the ( ___
- actual program is named, should have been supplied long ago __.=|___E
- as another one of those cryptic but standard Unix tools. /
- In the simplest usage, "nc host port" creates a TCP connection to the given
- port on the given target host. Your standard input is then sent to the host,
- and anything that comes back across the connection is sent to your standard
- output. This continues indefinitely, until the network side of the connection
- shuts down. Note that this behavior is different from most other applications
- which shut everything down and exit after an end-of-file on the standard input.
- Netcat can also function as a server, by listening for inbound connections
- on arbitrary ports and then doing the same reading and writing. With minor
- limitations, netcat doesn't really care if it runs in "client" or "server"
- mode -- it still shovels data back and forth until there isn't any more left.
- In either mode, shutdown can be forced after a configurable time of inactivity
- on the network side.
- And it can do this via UDP too, so netcat is possibly the "udp telnet-like"
- application you always wanted for testing your UDP-mode servers. UDP, as the
- "U" implies, gives less reliable data transmission than TCP connections and
- some systems may have trouble sending large amounts of data that way, but it's
- still a useful capability to have.
- You may be asking "why not just use telnet to connect to arbitrary ports?"
- Valid question, and here are some reasons. Telnet has the "standard input
- EOF" problem, so one must introduce calculated delays in driving scripts to
- allow network output to finish. This is the main reason netcat stays running
- until the *network* side closes. Telnet also will not transfer arbitrary
- binary data, because certain characters are interpreted as telnet options and
- are thus removed from the data stream. Telnet also emits some of its
- diagnostic messages to standard output, where netcat keeps such things
- religiously separated from its *output* and will never modify any of the real
- data in transit unless you *really* want it to. And of course telnet is
- incapable of listening for inbound connections, or using UDP instead. Netcat
- doesn't have any of these limitations, is much smaller and faster than telnet,
- and has many other advantages.
- Some of netcat's major features are:
- Outbound or inbound connections, TCP or UDP, to or from any ports
- Full DNS forward/reverse checking, with appropriate warnings
- Ability to use any local source port
- Ability to use any locally-configured network source address
- Built-in port-scanning capabilities, with randomizer
- Built-in loose source-routing capability
- Can read command line arguments from standard input
- Slow-send mode, one line every N seconds
- Hex dump of transmitted and received data
- Optional ability to let another program service established connections
- Optional telnet-options responder
- Efforts have been made to have netcat "do the right thing" in all its various
- modes. If you believe that it is doing the wrong thing under whatever
- circumstances, please notify me and tell me how you think it should behave.
- If netcat is not able to do some task you think up, minor tweaks to the code
- will probably fix that. It provides a basic and easily-modified template for
- writing other network applications, and I certainly encourage people to make
- custom mods and send in any improvements they make to it. This is the second
- release; the overall differences from 1.00 are relatively minor and have mostly
- to do with portability and bugfixes. Many people provided greatly appreciated
- fixes and comments on the 1.00 release. Continued feedback from the Internet
- community is always welcome!
- Netcat is entirely my own creation, although plenty of other code was used as
- examples. It is freely given away to the Internet community in the hope that
- it will be useful, with no restrictions except giving credit where it is due.
- No GPLs, Berkeley copyrights or any of that nonsense. The author assumes NO
- responsibility for how anyone uses it. If netcat makes you rich somehow and
- you're feeling generous, mail me a check. If you are affiliated in any way
- with Microsoft Network, get a life. Always ski in control. Comments,
- questions, and patches to hobbit@avian.org.
- Building
- ========
- Compiling is fairly straightforward. Examine the Makefile for a SYSTYPE that
- matches yours, and do "make <systype>". The executable "nc" should appear.
- If there is no relevant SYSTYPE section, try "generic". If you create new
- sections for generic.h and Makefile to support another platform, please follow
- the given format and mail back the diffs.
- There are a couple of other settable #defines in netcat.c, which you can
- include as DFLAGS="-DTHIS -DTHAT" to your "make" invocation without having to
- edit the Makefile. See the following discussions for what they are and do.
- If you want to link against the resolver library on SunOS [recommended] and
- you have BIND 4.9.x, you may need to change XLIBS=-lresolv in the Makefile to
- XLIBS="-lresolv -l44bsd".
- Linux sys/time.h does not really support presetting of FD_SETSIZE; a harmless
- warning is issued.
- Some systems may warn about pointer types for signal(). No problem, though.
- Exploration of features
- =======================
- Where to begin? Netcat is at the same time so simple and versatile, it's like
- trying to describe everything you can do with your Swiss Army knife. This will
- go over the basics; you should also read the usage examples and notes later on
- which may give you even more ideas about what this sort of tool is good for.
- If no command arguments are given at all, netcat asks for them, reads a line
- from standard input, and breaks it up into arguments internally. This can be
- useful when driving netcat from certain types of scripts, with the side effect
- of hiding your command line arguments from "ps" displays.
- The host argument can be a name or IP address. If -n is specified, netcat
- will only accept numeric IP addresses and do no DNS lookups for anything. If
- -n is not given and -v is turned on, netcat will do a full forward and reverse
- name and address lookup for the host, and warn you about the all-too-common
- problem of mismatched names in the DNS. This often takes a little longer for
- connection setup, but is useful to know about. There are circumstances under
- which this can *save* time, such as when you want to know the name for some IP
- address and also connect there. Netcat will just tell you all about it, saving
- the manual steps of looking up the hostname yourself. Normally mismatch-
- checking is case-insensitive per the DNS spec, but you can define ANAL at
- compile time to make it case-sensitive -- sometimes useful for uncovering minor
- errors in your own DNS files while poking around your networks.
- A port argument is required for outbound connections, and can be numeric or a
- name as listed in /etc/services. If -n is specified, only numeric arguments
- are valid. Special syntax and/or more than one port argument cause different
- behavior -- see details below about port-scanning.
- The -v switch controls the verbosity level of messages sent to standard error.
- You will probably want to run netcat most of the time with -v turned on, so you
- can see info about the connections it is trying to make. You will probably
- also want to give a smallish -w argument, which limits the time spent trying to
- make a connection. I usually alias "nc" to "nc -v -w 3", which makes it
- function just about the same for things I would otherwise use telnet to do.
- The timeout is easily changed by a subsequent -w argument which overrides the
- earlier one. Specifying -v more than once makes diagnostic output MORE
- verbose. If -v is not specified at all, netcat silently does its work unless
- some error happens, whereupon it describes the error and exits with a nonzero
- status. Refused network connections are generally NOT considered to be errors,
- unless you only asked for a single TCP port and it was refused.
- Note that -w also sets the network inactivity timeout. This does not have any
- effect until standard input closes, but then if nothing further arrives from
- the network in the next <timeout> seconds, netcat tries to read the net once
- more for good measure, and then closes and exits. There are a lot of network
- services now that accept a small amount of input and return a large amount of
- output, such as Gopher and Web servers, which is the main reason netcat was
- written to "block" on the network staying open rather than standard input.
- Handling the timeout this way gives uniform behavior with network servers that
- *don't* close by themselves until told to.
- UDP connections are opened instead of TCP when -u is specified. These aren't
- really "connections" per se since UDP is a connectionless protocol, although
- netcat does internally use the "connected UDP socket" mechanism that most
- kernels support. Although netcat claims that an outgoing UDP connection is
- "open" immediately, no data is sent until something is read from standard
- input. Only thereafter is it possible to determine whether there really is a
- UDP server on the other end, and often you just can't tell. Most UDP protocols
- use timeouts and retries to do their thing and in many cases won't bother
- answering at all, so you should specify a timeout and hope for the best. You
- will get more out of UDP connections if standard input is fed from a source
- of data that looks like various kinds of server requests.
- To obtain a hex dump file of the data sent either way, use "-o logfile". The
- dump lines begin with "<" or ">" to respectively indicate "from the net" or
- "to the net", and contain the total count per direction, and hex and ascii
- representations of the traffic. Capturing a hex dump naturally slows netcat
- down a bit, so don't use it where speed is critical.
- Netcat can bind to any local port, subject to privilege restrictions and ports
- that are already in use. It is also possible to use a specific local network
- source address if it is that of a network interface on your machine. [Note:
- this does not work correctly on all platforms.] Use "-p portarg" to grab a
- specific local port, and "-s ip-addr" or "-s name" to have that be your source
- IP address. This is often referred to as "anchoring the socket". Root users
- can grab any unused source port including the "reserved" ones less than 1024.
- Absence of -p will bind to whatever unused port the system gives you, just like
- any other normal client connection, unless you use -r [see below].
- Listen mode will cause netcat to wait for an inbound connection, and then the
- same data transfer happens. Thus, you can do "nc -l -p 1234 < filename" and
- when someone else connects to your port 1234, the file is sent to them whether
- they wanted it or not. Listen mode is generally used along with a local port
- argument -- this is required for UDP mode, while TCP mode can have the system
- assign one and tell you what it is if -v is turned on. If you specify a target
- host and optional port in listen mode, netcat will accept an inbound connection
- only from that host and if you specify one, only from that foreign source port.
- In verbose mode you'll be informed about the inbound connection, including what
- address and port it came from, and since listening on "any" applies to several
- possibilities, which address it came *to* on your end. If the system supports
- IP socket options, netcat will attempt to retrieve any such options from an
- inbound connection and print them out in hex.
- If netcat is compiled with -DGAPING_SECURITY_HOLE, the -e argument specifies
- a program to exec after making or receiving a successful connection. In the
- listening mode, this works similarly to "inetd" but only for a single instance.
- Use with GREAT CARE. This piece of the code is normally not enabled; if you
- know what you're doing, have fun. This hack also works in UDP mode. Note that
- you can only supply -e with the name of the program, but no arguments. If you
- want to launch something with an argument list, write a two-line wrapper script
- or just use inetd like always.
- If netcat is compiled with -DTELNET, the -t argument enables it to respond
- to telnet option negotiation [always in the negative, i.e. DONT or WONT].
- This allows it to connect to a telnetd and get past the initial negotiation
- far enough to get a login prompt from the server. Since this feature has
- the potential to modify the data stream, it is not enabled by default. You
- have to understand why you might need this and turn on the #define yourself.
- Data from the network connection is always delivered to standard output as
- efficiently as possible, using large 8K reads and writes. Standard input is
- normally sent to the net the same way, but the -i switch specifies an "interval
- time" which slows this down considerably. Standard input is still read in
- large batches, but netcat then tries to find where line breaks exist and sends
- one line every interval time. Note that if standard input is a terminal, data
- is already read line by line, so unless you make the -i interval rather long,
- what you type will go out at a fairly normal rate. -i is really designed
- for use when you want to "measure out" what is read from files or pipes.
- Port-scanning is a popular method for exploring what's out there. Netcat
- accepts its commands with options first, then the target host, and everything
- thereafter is interpreted as port names or numbers, or ranges of ports in M-N
- syntax. CAVEAT: some port names in /etc/services contain hyphens -- netcat
- currently will not correctly parse those, so specify ranges using numbers if
- you can. If more than one port is thus specified, netcat connects to *all* of
- them, sending the same batch of data from standard input [up to 8K worth] to
- each one that is successfully connected to. Specifying multiple ports also
- suppresses diagnostic messages about refused connections, unless -v is
- specified twice for "more verbosity". This way you normally get notified only
- about genuinely open connections. Example: "nc -v -w 2 -z target 20-30" will
- try connecting to every port between 20 and 30 [inclusive] at the target, and
- will likely inform you about an FTP server, telnet server, and mailer along the
- way. The -z switch prevents sending any data to a TCP connection and very
- limited probe data to a UDP connection, and is thus useful as a fast scanning
- mode just to see what ports the target is listening on. To limit scanning
- speed if desired, -i will insert a delay between each port probe. There are
- some pitfalls with regard to UDP scanning, described later, but in general it
- works well.
- For each range of ports specified, scanning is normally done downward within
- that range. If the -r switch is used, scanning hops randomly around within
- that range and reports open ports as it finds them. [If you want them listed
- in order regardless, pipe standard error through "sort"...] In addition, if
- random mode is in effect, the local source ports are also randomized. This
- prevents netcat from exhibiting any kind of regular pattern in its scanning.
- You can exert fairly fine control over your scan by judicious use of -r and
- selected port ranges to cover. If you use -r for a single connection, the
- source port will have a random value above 8192, rather than the next one the
- kernel would have assigned you. Note that selecting a specific local port
- with -p overrides any local-port randomization.
- Many people are interested in testing network connectivity using IP source
- routing, even if it's only to make sure their own firewalls are blocking
- source-routed packets. On systems that support it, the -g switch can be used
- multiple times [up to 8] to construct a loose-source-routed path for your
- connection, and the -G argument positions the "hop pointer" within the list.
- If your network allows source-routed traffic in and out, you can test
- connectivity to your own services via remote points in the internet. Note that
- although newer BSD-flavor telnets also have source-routing capability, it isn't
- clearly documented and the command syntax is somewhat clumsy. Netcat's
- handling of "-g" is modeled after "traceroute".
- Netcat tries its best to behave just like "cat". It currently does nothing to
- terminal input modes, and does no end-of-line conversion. Standard input from
- a terminal is read line by line with normal editing characters in effect. You
- can freely suspend out of an interactive connection and resume. ^C or whatever
- your interrupt character is will make netcat close the network connection and
- exit. A switch to place the terminal in raw mode has been considered, but so
- far has not been necessary. You can send raw binary data by reading it out of
- a file or piping from another program, so more meaningful effort would be spent
- writing an appropriate front-end driver.
- Netcat is not an "arbitrary packet generator", but the ability to talk to raw
- sockets and/or nit/bpf/dlpi may appear at some point. Such things are clearly
- useful; I refer you to Darren Reed's excellent ip_filter package, which now
- includes a tool to construct and send raw packets with any contents you want.
- Example uses -- the light side
- ==============================
- Again, this is a very partial list of possibilities, but it may get you to
- think up more applications for netcat. Driving netcat with simple shell or
- expect scripts is an easy and flexible way to do fairly complex tasks,
- especially if you're not into coding network tools in C. My coding isn't
- particularly strong either [although undoubtedly better after writing this
- thing!], so I tend to construct bare-metal tools like this that I can trivially
- plug into other applications. Netcat doubles as a teaching tool -- one can
- learn a great deal about more complex network protocols by trying to simulate
- them through raw connections!
- An example of netcat as a backend for something else is the shell-script
- Web browser, which simply asks for the relevant parts of a URL and pipes
- "GET /what/ever" into a netcat connection to the server. I used to do this
- with telnet, and had to use calculated sleep times and other stupidity to
- kludge around telnet's limitations. Netcat guarantees that I get the whole
- page, and since it transfers all the data unmodified, I can even pull down
- binary image files and display them elsewhere later. Some folks may find the
- idea of a shell-script web browser silly and strange, but it starts up and
- gets me my info a hell of a lot faster than a GUI browser and doesn't hide
- any contents of links and forms and such. This is included, as scripts/web,
- along with several other web-related examples.
- Netcat is an obvious replacement for telnet as a tool for talking to daemons.
- For example, it is easier to type "nc host 25", talk to someone's mailer, and
- just ^C out than having to type ^]c or QUIT as telnet would require you to do.
- You can quickly catalog the services on your network by telling netcat to
- connect to well-known services and collect greetings, or at least scan for open
- ports. You'll probably want to collect netcat's diagnostic messages in your
- output files, so be sure to include standard error in the output using
- `>& file' in *csh or `> file 2>&1' in bourne shell.
- A scanning example: "echo QUIT | nc -v -w 5 target 20-250 500-600 5990-7000"
- will inform you about a target's various well-known TCP servers, including
- r-services, X, IRC, and maybe a few you didn't expect. Sending in QUIT and
- using the timeout will almost guarantee that you see some kind of greeting or
- error from each service, which usually indicates what it is and what version.
- [Beware of the "chargen" port, though...] SATAN uses exactly this technique to
- collect host information, and indeed some of the ideas herein were taken from
- the SATAN backend tools. If you script this up to try every host in your
- subnet space and just let it run, you will not only see all the services,
- you'll find out about hosts that aren't correctly listed in your DNS. Then you
- can compare new snapshots against old snapshots to see changes. For going
- after particular services, a more intrusive example is in scripts/probe.
- Netcat can be used as a simple data transfer agent, and it doesn't really
- matter which end is the listener and which end is the client -- input at one
- side arrives at the other side as output. It is helpful to start the listener
- at the receiving side with no timeout specified, and then give the sending side
- a small timeout. That way the listener stays listening until you contact it,
- and after data stops flowing the client will time out, shut down, and take the
- listener with it. Unless the intervening network is fraught with problems,
- this should be completely reliable, and you can always increase the timeout. A
- typical example of something "rsh" is often used for: on one side,
- nc -l -p 1234 | uncompress -c | tar xvfp -
- and then on the other side
- tar cfp - /some/dir | compress -c | nc -w 3 othermachine 1234
- will transfer the contents of a directory from one machine to another, without
- having to worry about .rhosts files, user accounts, or inetd configurations
- at either end. Again, it matters not which is the listener or receiver; the
- "tarring" machine could just as easily be running the listener instead. One
- could conceivably use a scheme like this for backups, by having cron-jobs fire
- up listeners and backup handlers [which can be restricted to specific addresses
- and ports between each other] and pipe "dump" or "tar" on one machine to "dd
- of=/dev/tapedrive" on another as usual. Since netcat returns a nonzero exit
- status for a denied listener connection, scripts to handle such tasks could
- easily log and reject connect attempts from third parties, and then retry.
- Another simple data-transfer example: shipping things to a PC that doesn't have
- any network applications yet except a TCP stack and a web browser. Point the
- browser at an arbitrary port on a Unix server by telling it to download
- something like http://unixbox:4444/foo, and have a listener on the Unix side
- ready to ship out a file when the connect comes in. The browser may pervert
- binary data when told to save the URL, but you can dig the raw data out of
- the on-disk cache.
- If you build netcat with GAPING_SECURITY_HOLE defined, you can use it as an
- "inetd" substitute to test experimental network servers that would otherwise
- run under "inetd". A script or program will have its input and output hooked
- to the network the same way, perhaps sans some fancier signal handling. Given
- that most network services do not bind to a particular local address, whether
- they are under "inetd" or not, it is possible for netcat avoid the "address
- already in use" error by binding to a specific address. This lets you [as
- root, for low ports] place netcat "in the way" of a standard service, since
- inbound connections are generally sent to such specifically-bound listeners
- first and fall back to the ones bound to "any". This allows for a one-off
- experimental simulation of some service, without having to screw around with
- inetd.conf. Running with -v turned on and collecting a connection log from
- standard error is recommended.
- Netcat as well can make an outbound connection and then run a program or script
- on the originating end, with input and output connected to the same network
- port. This "inverse inetd" capability could enhance the backup-server concept
- described above or help facilitate things such as a "network dialback" concept.
- The possibilities are many and varied here; if such things are intended as
- security mechanisms, it may be best to modify netcat specifically for the
- purpose instead of wrapping such functions in scripts.
- Speaking of inetd, netcat will function perfectly well *under* inetd as a TCP
- connection redirector for inbound services, like a "plug-gw" without the
- authentication step. This is very useful for doing stuff like redirecting
- traffic through your firewall out to other places like web servers and mail
- hubs, while posing no risk to the firewall machine itself. Put netcat behind
- inetd and tcp_wrappers, perhaps thusly:
- www stream tcp nowait nobody /etc/tcpd /bin/nc -w 3 realwww 80
- and you have a simple and effective "application relay" with access control
- and logging. Note use of the wait time as a "safety" in case realwww isn't
- reachable or the calling user aborts the connection -- otherwise the relay may
- hang there forever.
- You can use netcat to generate huge amounts of useless network data for
- various performance testing. For example, doing
- yes AAAAAAAAAAAAAAAAAAAAAA | nc -v -v -l -p 2222 > /dev/null
- on one side and then hitting it with
- yes BBBBBBBBBBBBBBBBBBBBBB | nc othermachine 2222 > /dev/null
- from another host will saturate your wires with A's and B's. The "very
- verbose" switch usage will tell you how many of each were sent and received
- after you interrupt either side. Using UDP mode produces tremendously MORE
- trash per unit time in the form of fragmented 8 Kbyte mobygrams -- enough to
- stress-test kernels and network interfaces. Firing random binary data into
- various network servers may help expose bugs in their input handling, which
- nowadays is a popular thing to explore. A simple example data-generator is
- given in data/data.c included in this package, along with a small collection
- of canned input files to generate various packet contents. This program is
- documented in its beginning comments, but of interest here is using "%r" to
- generate random bytes at well-chosen points in a data stream. If you can
- crash your daemon, you likely have a security problem.
- The hex dump feature may be useful for debugging odd network protocols,
- especially if you don't have any network monitoring equipment handy or aren't
- root where you'd need to run "tcpdump" or something. Bind a listening netcat
- to a local port, and have it run a script which in turn runs another netcat
- to the real service and captures the hex dump to a log file. This sets up a
- transparent relay between your local port and wherever the real service is.
- Be sure that the script-run netcat does *not* use -v, or the extra info it
- sends to standard error may confuse the protocol. Note also that you cannot
- have the "listen/exec" netcat do the data capture, since once the connection
- arrives it is no longer netcat that is running.
- Binding to an arbitrary local port allows you to simulate things like r-service
- clients, if you are root locally. For example, feeding "^@root^@joe^@pwd^@"
- [where ^@ is a null, and root/joe could be any other local/remote username
- pair] into a "rsh" or "rlogin" server, FROM your port 1023 for example,
- duplicates what the server expects to receive. Thus, you can test for insecure
- .rhosts files around your network without having to create new user accounts on
- your client machine. The program data/rservice.c can aid this process by
- constructing the "rcmd" protocol bytes. Doing this also prevents "rshd" from
- trying to create that separate standard-error socket and still gives you an
- input path, as opposed to the usual action of "rsh -n". Using netcat for
- things like this can be really useful sometimes, because rsh and rlogin
- generally want a host *name* as an argument and won't accept IP addresses. If
- your client-end DNS is hosed, as may be true when you're trying to extract
- backup sets on to a dumb client, "netcat -n" wins where normal rsh/rlogin is
- useless.
- If you are unsure that a remote syslogger is working, test it with netcat.
- Make a UDP connection to port 514 and type in "<0>message", which should
- correspond to "kern.emerg" and cause syslogd to scream into every file it has
- open [and possibly all over users' terminals]. You can tame this down by
- using a different number and use netcat inside routine scripts to send syslog
- messages to places that aren't configured in syslog.conf. For example,
- "echo '<38>message' | nc -w 1 -u loggerhost 514" should send to auth.notice
- on loggerhost. The exact number may vary; check against your syslog.h first.
- Netcat provides several ways for you to test your own packet filters. If you
- bind to a port normally protected against outside access and make a connection
- to somewhere outside your own network, the return traffic will be coming to
- your chosen port from the "outside" and should be blocked. TCP may get through
- if your filter passes all "ack syn", but it shouldn't be even doing that to low
- ports on your network. Remember to test with UDP traffic as well! If your
- filter passes at least outbound source-routed IP packets, bouncing a connection
- back to yourself via some gateway outside your network will create "incoming"
- traffic with your source address, which should get dropped by a correctly
- configured anti-spoofing filter. This is a "non-test" if you're also dropping
- source-routing, but it's good to be able to test for that too. Any packet
- filter worth its salt will be blocking source-routed packets in both
- directions, but you never know what interesting quirks you might turn up by
- playing around with source ports and addresses and watching the wires with a
- network monitor.
- You can use netcat to protect your own workstation's X server against outside
- access. X is stupid enough to listen for connections on "any" and never tell
- you when new connections arrive, which is one reason it is so vulnerable. Once
- you have all your various X windows up and running you can use netcat to bind
- just to your ethernet address and listen to port 6000. Any new connections
- from outside the machine will hit netcat instead your X server, and you get a
- log of who's trying. You can either tell netcat to drop the connection, or
- perhaps run another copy of itself to relay to your actual X server on
- "localhost". This may not work for dedicated X terminals, but it may be
- possible to authorize your X terminal only for its boot server, and run a relay
- netcat over on the server that will in turn talk to your X terminal. Since
- netcat only handles one listening connection per run, make sure that whatever
- way you rig it causes another one to run and listen on 6000 soon afterward, or
- your real X server will be reachable once again. A very minimal script just
- to protect yourself could be
- while true ; do
- nc -v -l -s <your-addr> -p 6000 localhost 2
- done
- which causes netcat to accept and then close any inbound connection to your
- workstation's normal ethernet address, and another copy is immediately run by
- the script. Send standard error to a file for a log of connection attempts.
- If your system can't do the "specific bind" thing all is not lost; run your
- X server on display ":1" or port 6001, and netcat can still function as a probe
- alarm by listening on 6000.
- Does your shell-account provider allow personal Web pages, but not CGI scripts?
- You can have netcat listen on a particular port to execute a program or script
- of your choosing, and then just point to the port with a URL in your homepage.
- The listener could even exist on a completely different machine, avoiding the
- potential ire of the homepage-host administrators. Since the script will get
- the raw browser query as input it won't look like a typical CGI script, and
- since it's running under your UID you need to write it carefully. You may want
- to write a netcat-based script as a wrapper that reads a query and sets up
- environment variables for a regular CGI script. The possibilities for using
- netcat and scripts to handle Web stuff are almost endless. Again, see the
- examples under scripts/.
- Example uses -- the dark side
- =============================
- Equal time is deserved here, since a versatile tool like this can be useful
- to any Shade of Hat. I could use my Victorinox to either fix your car or
- disassemble it, right? You can clearly use something like netcat to attack
- or defend -- I don't try to govern anyone's social outlook, I just build tools.
- Regardless of your intentions, you should still be aware of these threats to
- your own systems.
- The first obvious thing is scanning someone *else's* network for vulnerable
- services. Files containing preconstructed data, be it exploratory or
- exploitive, can be fed in as standard input, including command-line arguments
- to netcat itself to keep "ps" ignorant of your doings. The more random the
- scanning, the less likelihood of detection by humans, scan-detectors, or
- dynamic filtering, and with -i you'll wait longer but avoid loading down the
- target's network. Some examples for crafting various standard UDP probes are
- given in data/*.d.
- Some configurations of packet filters attempt to solve the FTP-data problem by
- just allowing such connections from the outside. These come FROM port 20, TO
- high TCP ports inside -- if you locally bind to port 20, you may find yourself
- able to bypass filtering in some cases. Maybe not to low ports "inside", but
- perhaps to TCP NFS servers, X servers, Prospero, ciscos that listen on 200x
- and 400x... Similar bypassing may be possible for UDP [and maybe TCP too] if a
- connection comes from port 53; a filter may assume it's a nameserver response.
- Using -e in conjunction with binding to a specific address can enable "server
- takeover" by getting in ahead of the real ones, whereupon you can snarf data
- sent in and feed your own back out. At the very least you can log a hex dump
- of someone else's session. If you are root, you can certainly use -s and -e to
- run various hacked daemons without having to touch inetd.conf or the real
- daemons themselves. You may not always have the root access to deal with low
- ports, but what if you are on a machine that also happens to be an NFS server?
- You might be able to collect some interesting things from port 2049, including
- local file handles. There are several other servers that run on high ports
- that are likely candidates for takeover, including many of the RPC services on
- some platforms [yppasswdd, anyone?]. Kerberos tickets, X cookies, and IRC
- traffic also come to mind. RADIUS-based terminal servers connect incoming
- users to shell-account machines on a high port, usually 1642 or thereabouts.
- SOCKS servers run on 1080. Do "netstat -a" and get creative.
- There are some daemons that are well-written enough to bind separately to all
- the local interfaces, possibly with an eye toward heading off this sort of
- problem. Named from recent BIND releases, and NTP, are two that come to mind.
- Netstat will show these listening on address.53 instead of *.53. You won't
- be able to get in front of these on any of the real interface addresses, which
- of course is especially interesting in the case of named, but these servers
- sometimes forget about things like "alias" interface addresses or interfaces
- that appear later on such as dynamic PPP links. There are some hacked web
- servers and versions of "inetd" floating around that specifically bind as well,
- based on a configuration file -- these generally *are* bound to alias addresses
- to offer several different address-based services from one machine.
- Using -e to start a remote backdoor shell is another obvious sort of thing,
- easier than constructing a file for inetd to listen on "ingreslock" or
- something, and you can access-control it against other people by specifying a
- client host and port. Experience with this truly demonstrates how fragile the
- barrier between being "logged in" or not really is, and is further expressed by
- scripts/bsh. If you're already behind a firewall, it may be easier to make an
- *outbound* connection and then run a shell; a small wrapper script can
- periodically try connecting to a known place and port, you can later listen
- there until the inbound connection arrives, and there's your shell. Running
- a shell via UDP has several interesting features, although be aware that once
- "connected", the UDP stub sockets tend to show up in "netstat" just like TCP
- connections and may not be quite as subtle as you wanted. Packets may also be
- lost, so use TCP if you need reliable connections. But since UDP is
- connectionless, a hookup of this sort will stick around almost forever, even if
- you ^C out of netcat or do a reboot on your side, and you only need to remember
- the ports you used on both ends to reestablish. And outbound UDP-plus-exec
- connection creates the connected socket and starts the program immediately. On
- a listening UDP connection, the socket is created once a first packet is
- received. In either case, though, such a "connection" has the interesting side
- effect that only your client-side IP address and [chosen?] source port will
- thereafter be able to talk to it. Instant access control! A non-local third
- party would have to do ALL of the following to take over such a session:
- forge UDP with your source address [trivial to do; see below]
- guess the port numbers of BOTH ends, or sniff the wire for them
- arrange to block ICMP or UDP return traffic between it and your real
- source, so the session doesn't die with a network write error.
- The companion program data/rservice.c is helpful in scripting up any sort of
- r-service username or password guessing attack. The arguments to "rservice"
- are simply the strings that get null-terminated and passed over an "rcmd"-style
- connection, with the assumption that the client does not need a separate
- standard-error port. Brute-force password banging is best done via "rexec" if
- it is available since it is less likely to log failed attempts. Thus, doing
- "rservice joe joespass pwd | nc target exec" should return joe's home dir if
- the password is right, or "Permission denied." Plug in a dictionary and go to
- town. If you're attacking rsh/rlogin, remember to be root and bind to a port
- between 512 and 1023 on your end, and pipe in "rservice joe joe pwd" and such.
- Netcat can prevent inadvertently sending extra information over a telnet
- connection. Use "nc -t" in place of telnet, and daemons that try to ask for
- things like USER and TERM environment variables will get no useful answers, as
- they otherwise would from a more recent telnet program. Some telnetds actually
- try to collect this stuff and then plug the USER variable into "login" so that
- the caller is then just asked for a password! This mechanism could cause a
- login attempt as YOUR real username to be logged over there if you use a
- Borman-based telnet instead of "nc -t".
- Got an unused network interface configured in your kernel [e.g. SLIP], or
- support for alias addresses? Ifconfig one to be any address you like, and bind
- to it with -s to enable all sorts of shenanigans with bogus source addresses.
- The interface probably has to be UP before this works; some SLIP versions
- need a far-end address before this is true. Hammering on UDP services is then
- a no-brainer. What you can do to an unfiltered syslog daemon should be fairly
- obvious; trimming the conf file can help protect against it. Many routers out
- there still blindly believe what they receive via RIP and other routing
- protocols. Although most UDP echo and chargen servers check if an incoming
- packet was sent from *another* "internal" UDP server, there are many that still
- do not, any two of which [or many, for that matter] could keep each other
- entertained for hours at the expense of bandwidth. And you can always make
- someone wonder why she's being probed by nsa.gov.
- Your TCP spoofing possibilities are mostly limited to destinations you can
- source-route to while locally bound to your phony address. Many sites block
- source-routed packets these days for precisely this reason. If your kernel
- does oddball things when sending source-routed packets, try moving the pointer
- around with -G. You may also have to fiddle with the routing on your own
- machine before you start receiving packets back. Warning: some machines still
- send out traffic using the source address of the outbound interface, regardless
- of your binding, especially in the case of localhost. Check first. If you can
- open a connection but then get no data back from it, the target host is
- probably killing the IP options on its end [this is an option inside TCP
- wrappers and several other packages], which happens after the 3-way handshake
- is completed. If you send some data and observe the "send-q" side of "netstat"
- for that connection increasing but never getting sent, that's another symptom.
- Beware: if Sendmail 8.7.x detects a source-routed SMTP connection, it extracts
- the hop list and sticks it in the Received: header!
- SYN bombing [sometimes called "hosing"] can disable many TCP servers, and if
- you hit one often enough, you can keep it unreachable for days. As is true of
- many other denial-of-service attacks, there is currently no defense against it
- except maybe at the human level. Making kernel SOMAXCONN considerably larger
- than the default and the half-open timeout smaller can help, and indeed some
- people running large high-performance web servers have *had* to do that just to
- handle normal traffic. Taking out mailers and web servers is sociopathic, but
- on the other hand it is sometimes useful to be able to, say, disable a site's
- identd daemon for a few minutes. If someone realizes what is going on,
- backtracing will still be difficult since the packets have a phony source
- address, but calls to enough ISP NOCs might eventually pinpoint the source.
- It is also trivial for a clueful ISP to watch for or even block outgoing
- packets with obviously fake source addresses, but as we know many of them are
- not clueful or willing to get involved in such hassles. Besides, outbound
- packets with an [otherwise unreachable] source address in one of their net
- blocks would look fairly legitimate.
- Notes
- =====
- A discussion of various caveats, subtleties, and the design of the innards.
- As of version 1.07 you can construct a single file containing command arguments
- and then some data to transfer. Netcat is now smart enough to pick out the
- first line and build the argument list, and send any remaining data across the
- net to one or multiple ports. The first release of netcat had trouble with
- this -- it called fgets() for the command line argument, which behind the
- scenes does a large read() from standard input, perhaps 4096 bytes or so, and
- feeds that out to the fgets() library routine. By the time netcat 1.00 started
- directly read()ing stdin for more data, 4096 bytes of it were gone. It now
- uses raw read() everywhere and does the right thing whether reading from files,
- pipes, or ttys. If you use this for multiple-port connections, the single
- block of data will now be a maximum of 8K minus the first line. Improvements
- have been made to the logic in sending the saved chunk to each new port. Note
- that any command-line arguments hidden using this mechanism could still be
- extracted from a core dump.
- When netcat receives an inbound UDP connection, it creates a "connected socket"
- back to the source of the connection so that it can also send out data using
- normal write(). Using this mechanism instead of recvfrom/sendto has several
- advantages -- the read/write select loop is simplified, and ICMP errors can in
- effect be received by non-root users. However, it has the subtle side effect
- that if further UDP packets arrive from the caller but from different source
- ports, the listener will not receive them. UDP listen mode on a multihomed
- machine may have similar quirks unless you specifically bind to one of its
- addresses. It is not clear that kernel support for UDP connected sockets
- and/or my understanding of it is entirely complete here, so experiment...
- You should be aware of some subtleties concerning UDP scanning. If -z is on,
- netcat attempts to send a single null byte to the target port, twice, with a
- small time in between. You can either use the -w timeout, or netcat will try
- to make a "sideline" TCP connection to the target to introduce a small time
- delay equal to the round-trip time between you and the target. Note that if
- you have a -w timeout and -i timeout set, BOTH take effect and you wait twice
- as long. The TCP connection is to a normally refused port to minimize traffic,
- but if you notice a UDP fast-scan taking somewhat longer than it should, it
- could be that the target is actually listening on the TCP port. Either way,
- any ICMP port-unreachable messages from the target should have arrived in the
- meantime. The second single-byte UDP probe is then sent. Under BSD kernels,
- the ICMP error is delivered to the "connected socket" and the second write
- returns an error, which tells netcat that there is NOT a UDP service there.
- While Linux seems to be a fortunate exception, under many SYSV derived kernels
- the ICMP is not delivered, and netcat starts reporting that *all* the ports are
- "open" -- clearly wrong. [Some systems may not even *have* the "udp connected
- socket" concept, and netcat in its current form will not work for UDP at all.]
- If -z is specified and only one UDP port is probed, netcat's exit status
- reflects whether the connection was "open" or "refused" as with TCP.
- It may also be that UDP packets are being blocked by filters with no ICMP error
- returns, in which case everything will time out and return "open". This all
- sounds backwards, but that's how UDP works. If you're not sure, try "echo
- w00gumz | nc -u -w 2 target 7" to see if you can reach its UDP echo port at
- all. You should have no trouble using a BSD-flavor system to scan for UDP
- around your own network, although flooding a target with the high activity that
- -z generates will cause it to occasionally drop packets and indicate false
- "opens". A more "correct" way to do this is collect and analyze the ICMP
- errors, as does SATAN's "udp_scan" backend, but then again there's no guarantee
- that the ICMP gets back to you either. Udp_scan also does the zero-byte
- probes but is excruciatingly careful to calculate its own round-trip timing
- average and dynamically set its own response timeouts along with decoding any
- ICMP received. Netcat uses a much sleazier method which is nonetheless quite
- effective. Cisco routers are known to have a "dead time" in between ICMP
- responses about unreachable UDP ports, so a fast scan of a cisco will show
- almost everything "open". If you are looking for a specific UDP service, you
- can construct a file containing the right bytes to trigger a response from the
- other end and send that as standard input. Netcat will read up to 8K of the
- file and send the same data to every UDP port given. Note that you must use a
- timeout in this case [as would any other UDP client application] since the
- two-write probe only happens if -z is specified.
- Many telnet servers insist on a specific set of option negotiations before
- presenting a login banner. On a raw connection you will see this as small
- amount of binary gook. My attempts to create fixed input bytes to make a
- telnetd happy worked some places but failed against newer BSD-flavor ones,
- possibly due to timing problems, but there are a couple of much better
- workarounds. First, compile with -DTELNET and use -t if you just want to get
- past the option negotiation and talk to something on a telnet port. You will
- still see the binary gook -- in fact you'll see a lot more of it as the options
- are responded to behind the scenes. The telnet responder does NOT update the
- total byte count, or show up in the hex dump -- it just responds negatively to
- any options read from the incoming data stream. If you want to use a normal
- full-blown telnet to get to something but also want some of netcat's features
- involved like settable ports or timeouts, construct a tiny "foo" script:
- #! /bin/sh
- exec nc -otheroptions targethost 23
- and then do
- nc -l -p someport -e foo localhost &
- telnet localhost someport
- and your telnet should connect transparently through the exec'ed netcat to
- the target, using whatever options you supplied in the "foo" script. Don't
- use -t inside the script, or you'll wind up sending *two* option responses.
- I've observed inconsistent behavior under some Linuxes [perhaps just older
- ones?] when binding in listen mode. Sometimes netcat binds only to "localhost"
- if invoked with no address or port arguments, and sometimes it is unable to
- bind to a specific address for listening if something else is already listening
- on "any". The former problem can be worked around by specifying "-s 0.0.0.0",
- which will do the right thing despite netcat claiming that it's listening on
- [127.0.0.1]. This is a known problem -- for example, there's a mention of it
- in the makefile for SOCKS. On the flip side, binding to localhost and sending
- packets to some other machine doesn't work as you'd expect -- they go out with
- the source address of the sending interface instead. The Linux kernel contains
- a specific check to ensure that packets from 127.0.0.1 are never sent to the
- wire; other kernels may contain similar code. Linux, of course, *still*
- doesn't support source-routing, but they claim that it and many other network
- improvements are at least breathing hard.
- There are several possible errors associated with making TCP connections, but
- to specifically see anything other than "refused", one must wait the full
- kernel-defined timeout for a connection to fail. Netcat's mechanism of
- wrapping an alarm timer around the connect prevents the *real* network error
- from being returned -- "errno" at that point indicates "interrupted system
- call" since the connect attempt was interrupted. Some old 4.3 BSD kernels
- would actually return things like "host unreachable" immediately if that was
- the case, but most newer kernels seem to wait the full timeout and *then* pass
- back the real error. Go figure. In this case, I'd argue that the old way was
- better, despite those same kernels generally being the ones that tear down
- *established* TCP connections when ICMP-bombed.
- Incoming socket options are passed to applications by the kernel in the
- kernel's own internal format. The socket-options structure for source-routing
- contains the "first-hop" IP address first, followed by the rest of the real
- options list. The kernel uses this as is when sending reply packets -- the
- structure is therefore designed to be more useful to the kernel than to humans,
- but the hex dump of it that netcat produces is still useful to have.
- Kernels treat source-routing options somewhat oddly, but it sort of makes sense
- once one understands what's going on internally. The options list of addresses
- must contain hop1, hop2, ..., destination. When a source-routed packet is sent
- by the kernel [at least BSD], the actual destination address becomes irrelevant
- because it is replaced with "hop1", "hop1" is removed from the options list,
- and all the other addresses in the list are shifted up to fill the hole. Thus
- the outbound packet is sent from your chosen source address to the first
- *gateway*, and the options list now contains hop2, ..., destination. During
- all this address shuffling, the kernel does NOT change the pointer value, which
- is why it is useful to be able to set the pointer yourself -- you can construct
- some really bizarre return paths, and send your traffic fairly directly to the
- target but around some larger loop on the way back. Some Sun kernels seem to
- never flip the source-route around if it contains less than three hops, never
- reset the pointer anyway, and tries to send the packet [with options containing
- a "completed" source route!!] directly back to the source. This is way broken,
- of course. [Maybe ipforwarding has to be on? I haven't had an opportunity to
- beat on it thoroughly yet.]
- "Credits" section: The original idea for netcat fell out of a long-standing
- desire and fruitless search for a tool resembling it and having the same
- features. After reading some other network code and realizing just how many
- cool things about sockets could be controlled by the calling user, I started
- on the basics and the rest fell together pretty quickly. Some port-scanning
- ideas were taken from Venema/Farmer's SATAN tool kit, and Pluvius' "pscan"
- utility. Healthy amounts of BSD kernel source were perused in an attempt to
- dope out socket options and source-route handling; additional help was obtained
- from Dave Borman's telnet sources. The select loop is loosely based on fairly
- well-known code from "rsh" and Richard Stevens' "sock" program [which itself is
- sort of a "netcat" with more obscure features], with some more paranoid
- sanity-checking thrown in to guard against the distinct likelihood that there
- are subtleties about such things I still don't understand. I found the
- argument-hiding method cleanly implemented in Barrett's "deslogin"; reading the
- line as input allows greater versatility and is much less prone to cause
- bizarre problems than the more common trick of overwriting the argv array.
- After the first release, several people contributed portability fixes; they are
- credited in generic.h and the Makefile. Lauren Burka inspired the ascii art
- for this revised document. Dean Gaudet at Wired supplied a precursor to
- the hex-dump code, and mudge@l0pht.com originally experimented with and
- supplied code for the telnet-options responder. Outbound "-e <prog>" resulted
- from a need to quietly bypass a firewall installation. Other suggestions and
- patches have rolled in for which I am always grateful, but there are only 26
- hours per day and a discussion of feature creep near the end of this document.
- Netcat was written with the Russian railroad in mind -- conservatively built
- and solid, but it *will* get you there. While the coding style is fairly
- "tight", I have attempted to present it cleanly [keeping *my* lines under 80
- characters, dammit] and put in plenty of comments as to why certain things
- are done. Items I know to be questionable are clearly marked with "XXX".
- Source code was made to be modified, but determining where to start is
- difficult with some of the tangles of spaghetti code that are out there.
- Here are some of the major points I feel are worth mentioning about netcat's
- internal design, whether or not you agree with my approach.
- Except for generic.h, which changes to adapt more platforms, netcat is a single
- source file. This has the distinct advantage of only having to include headers
- once and not having to re-declare all my functions in a billion different
- places. I have attempted to contain all the gross who's-got-what-.h-file
- things in one small dumping ground. Functions are placed "dependencies-first",
- such that when the compiler runs into the calls later, it already knows the
- type and arguments and won't complain. No function prototyping -- not even the
- __P(()) crock -- is used, since it is more portable and a file of this size is
- easy enough to check manually. Each function has a standard-format comment
- ahead of it, which is easily found using the regexp " :$". I freely use gotos.
- Loops and if-clauses are made as small and non-nested as possible, and the ends
- of same *marked* for clarity [I wish everyone would do this!!].
- Large structures and buffers are all malloc()ed up on the fly, slightly larger
- than the size asked for and zeroed out. This reduces the chances of damage
- from those "end of the buffer" fencepost errors or runaway pointers escaping
- off the end. These things are permanent per run, so nothing needs to be freed
- until the program exits.
- File descriptor zero is always expected to be standard input, even if it is
- closed. If a new network descriptor winds up being zero, a different one is
- asked for which will be nonzero, and fd zero is simply left kicking around
- for the rest of the run. Why? Because everything else assumes that stdin is
- always zero and "netfd" is always positive. This may seem silly, but it was a
- lot easier to code. The new fd is obtained directly as a new socket, because
- trying to simply dup() a new fd broke subsequent socket-style use of the new fd
- under Solaris' stupid streams handling in the socket library.
- The catch-all message and error handlers are implemented with an ample list of
- phoney arguments to get around various problems with varargs. Varargs seems
- like deliberate obfuscation in the first place, and using it would also
- require use of vfprintf() which not all platforms support. The trailing
- sleep in bail() is to allow output to flush, which is sometimes needed if
- netcat is already on the other end of a network connection.
- The reader may notice that the section that does DNS lookups seems much
- gnarlier and more confusing than other parts. This is NOT MY FAULT. The
- sockaddr and hostent abstractions are an abortion that forces the coder to
- deal with it. Then again, a lot of BSD kernel code looks like similar
- struct-pointer hell. I try to straighten it out somewhat by defining my own
- HINF structure, containing names, ascii-format IP addresses, and binary IP
- addresses. I fill this structure exactly once per host argument, and squirrel
- everything safely away and handy for whatever wants to reference it later.
- Where many other network apps use the FIONBIO ioctl to set non-blocking I/O
- on network sockets, netcat uses straightforward blocking I/O everywhere.
- This makes everything very lock-step, relying on the network and filesystem
- layers to feed in data when needed. Data read in is completely written out
- before any more is fetched. This may not be quite the right thing to do under
- some OSes that don't do timed select() right, but this remains to be seen.
- The hexdump routine is written to be as fast as possible, which is why it does
- so much work itself instead of just sprintf()ing everything together. Each
- dump line is built into a single buffer and atomically written out using the
- lowest level I/O calls. Further improvements could undoubtedly be made by
- using writev() and eliminating all sprintf()s, but it seems to fly right along
- as is. If both exec-a-prog mode and a hexdump file is asked for, the hexdump
- flag is deliberately turned off to avoid creating random zero-length files.
- Files are opened in "truncate" mode; if you want "append" mode instead, change
- the open flags in main().
- main() may look a bit hairy, but that's only because it has to go down the
- argv list and handle multiple ports, random mode, and exit status. Efforts
- have been made to place a minimum of code inside the getopt() loop. Any real
- work is sent off to functions in what is hopefully a straightforward way.
- Obligatory vendor-bash: If "nc" had become a standard utility years ago,
- the commercial vendors would have likely packaged it setuid root and with
- -DGAPING_SECURITY_HOLE turned on but not documented. It is hoped that netcat
- will aid people in finding and fixing the no-brainer holes of this sort that
- keep appearing, by allowing easier experimentation with the "bare metal" of
- the network layer.
- It could be argued that netcat already has too many features. I have tried
- to avoid "feature creep" by limiting netcat's base functionality only to those
- things which are truly relevant to making network connections and the everyday
- associated DNS lossage we're used to. Option switches already have slightly
- overloaded functionality. Random port mode is sort of pushing it. The
- hex-dump feature went in later because it *is* genuinely useful. The
- telnet-responder code *almost* verges on the gratuitous, especially since it
- mucks with the data stream, and is left as an optional piece. Many people have
- asked for example "how 'bout adding encryption?" and my response is that such
- things should be separate entities that could pipe their data *through* netcat
- instead of having their own networking code. I am therefore not completely
- enthusiastic about adding any more features to this thing, although you are
- still free to send along any mods you think are useful.
- Nonetheless, at this point I think of netcat as my tcp/ip swiss army knife,
- and the numerous companion programs and scripts to go with it as duct tape.
- Duct tape of course has a light side and a dark side and binds the universe
- together, and if I wrap enough of it around what I'm trying to accomplish,
- it *will* work. Alternatively, if netcat is a large hammer, there are many
- network protocols that are increasingly looking like nails by now...
- _H* 960320 v1.10 RELEASE -- happy spring!